Starship Build Engineering System
Verify
VERIFY — Starship Build Engineering System
How verification runs
node verify.mjs imports the engines directly from src/, runs 18 assertions, and writes verification-report.json (machine) and verification-report.md (human) plus a copy under evidence/. Exit code is non-zero if any check fails.
What each check asserts
| # | Check | What it proves |
|---|---|---|
| 1 | Takt = 480/8 = 60 | takt formula |
| 2 | Theoretical min stations = ceil(500/60) = 9 | work-content / takt bound |
| 3 | 4×30-min chain @ takt 60 → 2 stations, 100% eff | balancer packs tasks optimally in a known case |
| 4 | Vehicle balance respects precedence, station ≤ takt, min ≤ N ≤ ops | no illegal assignment |
| 5 | Line-efficiency identity = work/(stations·cycle), 0 < eff ≤ 1 | metric is internally consistent |
| 6 | Balancing is deterministic | same input → same layout |
| 7 | Throughput = 480/60 = 8 | bottleneck-limited rate |
| 8 | Parallel stations for 130 min @ takt 60 = 3 | parallelization sizing |
| 9 | CPM A(3),B(4)→C(2) → 6, critical B,C | forward/backward pass on a tiny known net |
| 10 | Vehicle flow → 56 h, critical A,B,E,F,G, slacks D=14 C=2 | full CPM on the model |
| 11 | Cp/Cpk of [9,10,11] vs 7..13 = 1.0/1.0 | capability indices |
| 12 | Φ(0)=0.5, Φ(1.96)≈0.975, symmetry, erf(0)≈0 | normal-CDF accuracy |
| 13 | ±3σ → yield ≈ 0.9973, ≈ 2700 PPM | fallout model |
| 14 | tol [0.1,0.2,0.2] → WC 0.5, RSS 0.3 | exact stack-up |
| 15 | RSS ≤ worst-case for every vehicle stack | statistical bound invariant |
| 16 | 40→12 parts = 70%, 100→40 s = 60% | DFM reduction math |
| 17 | DFA: 2 essential parts, 2 candidates | consolidation logic |
| 18 | Balanced line ≤ stations and ≥ efficiency vs baseline | benchmark sanity |
Checks 1–17 are data-independent (they use fixtures with known answers), so they keep guarding the math if you swap the notional model for real data. Check 18 is the synthetic benchmark on the model.
What verification does NOT assert
It does not assert real-world accuracy, real production rates, or that any real weld/fit/part figure is correct — the model is notional (see proof/LIMITATIONS.md).